TypiCore: A Hybrid Active Query Strategy for Class-Incremental Learning on Time Series

In real-world applications like industrial monitoring and healthcare, machine learning models must adapt to new classes over time (Class-Incremental Learning, or CIL). However, most existing CIL methods assume that all incoming data is fully labeled, which is rarely true in practice due to high annotation costs. This paper investigates Active Class-Incremental Learning (ACIL) for multivariate time series, aiming to identify query strategies that maximize model performance under strict labeling budgets.

The authors first conducted a systematic evaluation of various active learning query strategies—including uncertainty-based, distribution-aware, and hybrid methods—within a rehearsal-based continual learning framework. They found that uncertainty-based methods often perform poorly in data-scarce scenarios, while distribution-aware methods (like Core-Set) and typicality-based methods (like TypiClust) offer better, yet incomplete, solutions for balancing model plasticity and stability.

To address these limitations, the authors introduced TypiCore, a novel hybrid strategy. TypiCore alternates between two selection criteria across active learning cycles:

1. Typicality-based selection: Identifies samples that are representative of their class density, ensuring the memory buffer contains high-quality, prototypical examples.
2. Diversity-based selection: Uses geometric coverage to ensure the buffer captures the breadth of the underlying data distribution.

Evaluated on the TSCIL benchmark, TypiCore consistently outperformed existing baselines. By intelligently selecting which samples to label, the method achieved performance levels matching or exceeding fully supervised continual learning models while using only a fraction of the total available labels. This work demonstrates that for time series data, a hybrid approach that balances representativeness and diversity is superior to relying on a single selection heuristic, providing a practical pathway for deploying continual learning in label-constrained environments.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.